Pemodelan Prediksi Kebakaran Hutan Multiregional Menggunakan Stacked Ensemble dan Explainable Artificial Intelligence

Kanedi, Fidela Jovita (2026) Pemodelan Prediksi Kebakaran Hutan Multiregional Menggunakan Stacked Ensemble dan Explainable Artificial Intelligence. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6026242016-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (16MB) | Request a copy

Abstract

Perubahan iklim memicu peningkatan frekuensi dan intensitas kebakaran hutan di wilayah dengan karakteristik ekologis kontras seperti Australia, Amerika Serikat, dan Indonesia. Penelitian ini mengembangkan kerangka pemodelan multiregional untuk klasifikasi tingkat kepercayaan hotspot kebakaran (Low, Nominal, High) yang mengintegrasikan stacked ensemble learning, kalibrasi probabilitas, dan analisis Explainable Artificial Intelligence (SHAP). Dataset dibangun dari hotspot VIIRS NASA FIRMS yang dilengkapi 27 fitur meteorologi, vegetasi, topografi, dan antropogenik dengan total 1.738.140 sampel. Arsitektur stacking terdiri dari Random Forest, XGBoost, dan LightGBM sebagai base learner serta Logistic Regression sebagai meta learner, diuji pada empat skenario evaluasi: Global Random Split, Global Temporal Split, Cross-Region, dan Per-Region. Hasil eksperimen menunjukkan stacking memperoleh accuracy 0,7934 dengan macro F1-score 0,5311 dan ROC-AUC 0,8297 pada Global Random Split. Kalibrasi sigmoid secara konsisten menurunkan Expected Calibration Error, seperti dari 0,0802 menjadi 0,0231 pada skenario tersebut, meskipun disertai trade-off pada sensitivitas kelas minoritas. Analisis SHAP mengungkap heterogenitas determinan antarwilayah: Indonesia didominasi fitur meteorologis, Australia oleh kombinasi temporal dan topografi, sedangkan Amerika Serikat menunjukkan pola intermediate. Temuan ini memberi fondasi metodologis untuk sistem peringatan dini kebakaran hutan yang adaptif dan transparan.
===========================================================================================================================================
Climate change has intensified the frequency and severity of wildfires across regions with contrasting ecological characteristics such as Australia, the United States, and Indonesia. This study develops a multiregional modeling framework to classify wildfire hotspot confidence levels (Low, Nominal, High) by integrating stacked ensemble learning, probability calibration, and Explainable Artificial Intelligence (SHAP) analysis. The dataset was constructed from NASA FIRMS VIIRS hotspots augmented with 27 meteorological, vegetation, topographic, and anthropogenic features, totaling 1,738,140 samples. The stacking architecture comprises Random Forest, XGBoost, and LightGBM as base learners with Logistic Regression as the meta learner, evaluated under four scenarios: Global Random Split, Global Temporal Split, Cross-Region, and Per-Region. Experimental results show the stacking model attained an accuracy of 0.7934, macro F1-score of 0.5311, and ROC-AUC of 0.8297 on the Global Random Split. Sigmoid calibration consistently reduced the Expected Calibration Error, for instance from 0.0802 to 0.0231 in that scenario, although accompanied by a trade-off in minority-class sensitivity. SHAP analysis revealed regional heterogeneity in dominant determinants: Indonesia is driven by meteorological features, Australia by temporal–topographic combinations, while the United States exhibits an intermediate pattern. These findings provide a methodological foundation for adaptive and transparent wildfire early-warning systems.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Kebakaran hutan, Prediksi kebakaran, pemodelan multiregional, Stacked ensemble, Explainable Artificial Intelligence, Kalibrasi probabilitas, Forest fire, hotspot confidence classification, multiregional modeling, stacked ensemble, Explainable Artificial Intelligence, probability calibration.
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Information and Communication Technology > Information Systems > 59101-(S2) Master Thesis
Depositing User:	Fidela Jovita Kanedi
Date Deposited:	28 Jul 2026 01:39
Last Modified:	28 Jul 2026 01:39
URI:	http://repository.its.ac.id/id/eprint/138076

Actions (login required)

View Item